← Back to list
Colossus (Invest Like the Best / Business Breakdowns)Podcast18 Mar 2025Source: joincolossus.comHost: Patrick O'Shaughnessy

Alex Wiltschko - Giving Computers A Sense Of Smell - [Invest Like the Best, EP.415]

In plain words

This is about Osmo, an AI company that gives computers a sense of smell and uses it to make perfumes. The founder thinks smell is the last sense to digitize, and AI is just starting to crack it, threatening the old fragrance industry. Key holdings: ①Generation (Osmo's own brand, cutting custom perfume delivery from 12-18 months to weeks); ②StockX (uses Osmo to detect fake sneakers in 20 seconds); ③traditional fragrance firms (stuck with 300-year-old model, 90% of clients get stock samples, but high margins and sticky customers).

AI SummaryAI-generated · may contain errors · verify against the original

Alex Wiltschko founded Osmo, a company dedicated to giving computers a sense of smell by teaching machines to "read" and "write" scents using AI technology. Its first commercial application, Generation, has disrupted the traditional perfume industry, dramatically accelerating the custom fragrance cr

~12 min full read · 7 sections
Deep Analysis

Quick Take

Guest: Alex Wiltschko, Founder and CEO of Osmo, former Google Brain researcher, with a neuroscience background (Harvard) and AI startup experience (two companies acquired, previously joined Twitter's deep learning team). Main thread: How Osmo digitizes smell through AI (reading + writing scents), and uses this platform to enter the fragrance industry (Generation brand) and broader commercial applications. Most impactful take: Alex Wiltschko believes that digitizing smell is the "last piece of the puzzle" — computers can already see, hear, and touch, but smell is the only modality not yet conquered, and Osmo is climbing the AI S-curve for this modality, currently at the "very leftmost, just taking off" stage.


Theme 1: Osmo's "Read then Write" Loop Is the Core of AI-Driven Olfaction Digitization

Alex Wiltschko emphasized that to achieve computer olfaction, both "read" and "write" capabilities are necessary, and "write" is the key to verifying whether "read" is correct.

  • Read (GCMS + AI): Osmo uses a gas chromatography-mass spectrometry (GCMS) instrument to break down odors into individual molecules. This device "already has the accumulation of 12 Nobel Prize-level technologies." Osmo's breakthrough lies in replacing its "brain" with proprietary AI software — "the hardware is already good, but the map (software) connecting different hardware pieces is completely missing." GCMS separates odors via the "molecular marathon" principle: lighter molecules finish the track first, are individually shattered and weighed by the mass spectrometer, and then the AI reconstructs the molecular identity like solving a Sudoku puzzle. Osmo has also fully automated this process: "the interpretation, which was traditionally done partly by software and partly by humans, we have now fully implemented with software."
  • Write (Scent Printer + Formula): The scent printer can recombine digitized odor formulas into physical scents. Osmo's first successful case — Plum 1.0 — involved "digitizing and remotely reconstructing" the scent of fresh summer black plums. Alex described the moment: "What I smelled was just a plum, almost surreal — I literally fell out of my chair."
  • Closed-Loop Driving Active Learning: "If you can read and write, you can create this virtuous cycle: every run generates new data. ... Then you can tilt the process so that the new scents created the next day teach the system more — that's active learning, and it's how AI systems get smart quickly." Currently, Osmo has a physical inventory of approximately 10,000–20,000 AI-designed molecules, each with a digital twin.
  • Validation Method: In double-blind trials, Osmo predicted the odors of hundreds of molecules it had never seen before. The result: the AI's prediction accuracy was higher than the average human panelist — "if you were to add one more person to the panel, you would rather ask our software what it smells like than train another human to smell the real thing." Alex regards this as "passing the smell Turing test."

Theme 2: Generation – Disrupting the Traditional Fragrance Industry with AI-Human Fusion

Alex Wiltschko believes that the fragrance industry's 300-year-old business model (12-18 months customization cycle, 90% of clients receiving "stock samples" rather than genuine customization) is exactly where Osmo enters, and the Generation brand is its "first vertical application."

Traditional Model vs. Generation Model

Dimension Traditional Fragrance Companies Generation (Osmo)
Customization Process Submit written brief → wait 3 months → receive stock samples → return → wait another 3 months → repeat for 12-18 months Chat-like interface (similar to ChatGPT) → AI instant analysis → rapid formulation output
Customization Rate Over 90% of clients receive non-genuine customized products Genuine customization every time
Delivery Cycle 12-18 months Reduced to weeks/months
Value Proposition Large clients, large scale, high barriers Opens access to creators who previously could not access custom scents (e.g., Instagram creators)

Technical Principle: From Text to a 300-Dimensional Scent Map

Alex describes Generation's workflow: Users describe their brand concept and desired scent direction (e.g., "the fresh air at the top of a sequoia forest"), and Osmo embeds the text into its 300-dimensional scent map ("not three-dimensional like RGB, but about 300 dimensions — which is why scent had to wait for AI to be digitized"). The system decodes these coordinates into a scent source file and displays where the scent falls on a map of the "top 100 mass-market best-selling fragrances." "What used to take months, we compressed into minutes."

Business Logic: Why Not Sell Software, but Build a Brand Instead?

This choice is backed by a deep market insight: Successful platforms typically enter markets where "there are already many active participants," but "there aren't many fragrance companies in the first place." Alex shared a similar case of Metropolis (a parking software company) — the parking industry is not a mature buyer of software, so they became a parking company. "If I could sell software, believe me, we tried... but we probably wouldn't be having this conversation." Therefore, Osmo chose to become a brand (Generation), serving end users directly, while retaining the platform capability.


Theme 3: The Commercial Appeal of Fragrance – High Margins, High Stickiness, Anti-Cyclical, but Can AI Reshape a Landscape "Unchanged for 300 Years"?

Alex Wiltschko argues that the fragrance industry inherently possesses the characteristics of a "good business," but traditional players face the risk of disruption in the age of AI.

Good Business Characteristics

  • Anti-cyclical: "If people aren't buying luxury perfume, they're buying hand soap. 90% of household products have a scent. When one category declines, another usually rises."
  • High profit margins: "These are essentially manufacturing companies, but with margins at the level of non-manufacturing firms – because a huge amount of proprietary know-how is embedded in the final product."
  • High customer stickiness: If a CPG company's product runs out of stock, "the first reaction is to re-order from the existing supplier, not to re-bid the contract." The industry standard repurchase rate is "far above 50%."

Risk: AI's S-Curve Has Only Climbed to the "Far Left" of the Olfactory Modality

Here, Alex makes a structural judgment: AI sits at different points on the S-curve across modalities. Text is "basically done" ("We passed the Turing test, the training data is just one internet, and we've used it all up"), images are similar, and video still needs time. But for smell: "We are pushing AI up the S-curve for smell, and right now we are at the far left – just starting to take off." This means that traditional fragrance companies, if they rely on old models, will be left behind by the acceleration driven by AI.


Theme 4: Commercialization's "Ladder Route" – Not Climbing the Peak Directly, but Finding a "Gentle Enough Slope"

Alex Wiltschko shared a key strategic philosophy: Osmo does not directly assault the cliff for the ultimate goal of "giving computers a sense of smell," but instead finds a route where "it can pause along the way to build a business."

  • Three-step approach: ① Read/write technology (already achieved) → ② Generation brand (being launched) → ③ Higher-level applications such as health detection (future).
  • "Never lose sight of the summit, but don't climb the cliff directly." Specifically, each step makes the company "harder to kill" rather than "making the probability of success more risky." Alex admitted: "I want to do this for my whole life, not just flip the company and sell it. It must survive."
  • Currently suppressed "second peak": The area Alex is most excited about but "must restrain himself from spending time on" is human health detection — "The breath and sweat we excrete contain signals expelled from blood and organs. Dogs can detect them. We believe computers can too." But he emphasizes that the Generation platform itself will provide the data foundation for this goal, because "the odor designs of fruits, flowers, and vegetables have substantial molecular overlap with human odors and product odors — if you do well in one odor domain, it automatically helps you enter other adjacent domains."

Theme 5: Risk – Mother Nature may "pull the rug out from under our feet"

Alex candidly lists the biggest uncertainties facing Osmo:

  • Scientific risk: "I am always worried that one day Mother Nature will show up and say 'you're done' – for example, 2025, 2026, or 2027, not the year we crack the next sensory mystery, and then we're stuck."
  • Engineering risk: The sensor needs to shrink from "the size of two shoeboxes" to a "portable" level (needs to shrink by another 4-8 times); the scent printer is currently "only half the size of this table."
  • Data risk: "The fuel to climb the S-curve is data." Osmo had to build its own data infrastructure – including an entire building of people labeling scents every day, sensors running 24/7, robots creating scents – because "external data is not AI-compatible."

Mentioned Positions

Target Guest Attitude Key Data
StockX Bullish (already partnered) Detects fake shoes via smell (true/false determination), results in 20 seconds; sensor is currently about the size of two shoeboxes
Generation (Osmo's own brand) Bullish (just launched) Reduces custom fragrance cycle from 12-18 months to weeks/months; targets creators who previously couldn't access scent (e.g., Instagram creators)
Traditional fragrance industry (unnamed companies) Risk warning (challenger) Business model unchanged for 300 years; 90% of clients receive "stock samples"; repurchase rate >50%; recession-resistant, high margins
Ritz-Carlton, Gramercy Park Hotel Neutral (example) Already use custom scents as brand identity

Judgments Worth Remembering

1. "Computer olfaction is the last piece of the puzzle" (Alex Wiltschko) — Computers can already see, hear, and touch, but smell has not yet been digitized. Osmo believes this is the last "human sensory modality" not yet penetrated by AI, and it is just beginning its ascent from the far left of its S-curve.

2. "300-dimensional map — smell must wait for AI to digitize" (Alex Wiltschko) — Smell is not a three-dimensional space like RGB, but approximately 300 dimensions, so traditional methods cannot handle it. Osmo uses graph neural networks (GNN) to map chemical structures into this high-dimensional space, achieving "predicting smell from molecular structure".

3. "AI predictions are more accurate than the average human evaluator" (Alex Wiltschko) — In double-blind tests, Osmo's AI predicted smells with higher accuracy than the average of a human panel. "If you need to add one more person to the panel, you'd rather ask the software."

4. "Don't climb the cliff directly; find a path with a gentle enough slope" (Alex Wiltschko) — To achieve the ultimate goal of "fully digitizing human olfaction", Osmo chose to first launch the Generation brand, build a sustainable business in the market, and accumulate data along the way. Each step makes the company "harder to kill" rather than riskier.

5. "The fragrance industry is essentially a manufacturing business, but its profit margins are at the level of non-manufacturing businesses" (Alex Wiltschko) — The reason is that "a large amount of proprietary know-how is embedded in the final product". Traditional players are safe, but AI may change this barrier.

6. "AI is at different positions on the S-curve for different modalities" (Alex Wiltschko) — Text is nearly complete ("there is only one internet, and it has been used up"), images are similar, video still needs time, and smell is just beginning. This means Osmo's space is much larger than the traditional industry appears.

7. "If you do well in one smell domain, it will automatically help you enter other adjacent domains" (Alex Wiltschko) — The molecular overlap between fruit/floral scent design and human health detection is high, so the Generation platform naturally lays the foundation for future health applications.

8. "Mother Nature may pull the rug out from under you" (Alex Wiltschko) — The biggest uncertainty is the science itself (whether 2025/2026/2027 will be the "year to crack the next layer of olfactory secrets"), as well as the physical miniaturization engineering challenges of sensors and printers.